Applying machine learning to associate clinical factors with malnutrition risk in peritoneal dialysis patients: an...

Malnutrition frequently complicates peritoneal dialysis (PD) and associates with adverse outcomes, underscoring the clinical importance of its timely identification. This study aimed to develop and internally validate a machine learning-based assessment model to identify PD patients currently at malnutrition risk who need nutritional intervention.

In this cross-sectional study, 144 PD patients were evaluated for malnutrition risk using the Patient-Generated Subjective Global Assessment (PG-SGA). A PG-SGA score ≥4 was prespecified to indicate malnutrition risk warranting dietitian-led intervention. Candidate predictors included demographic characteristics, laboratory indices, and physical function measures. To avoid overfitting and selection bias associated with comparing multiple algorithms in a modest sample, we pre-specified three representative models: logistic regression (LR) as a linear baseline, penalized logistic regression (ridge), and extreme gradient boosting (XGBoost). Feature selection was performed using LASSO regression. Instead of a single training-validation split, we employed 1,000 bootstrap resampling iterations for internal validation. Model performance was assessed using the C-index, Brier score, calibration intercept, and calibration slope, with bootstrap-derived 95% confidence intervals.

Malnutrition risk was observed in 47.2% of participants (68/144). LASSO retained seven associated features: age, short physical performance battery (SPPB) score, timed up and go (TUG) test result, triceps skinfold thickness, handgrip strength, triglycerides, and serum albumin. Based on bootstrap internal validation, the XGBoost model achieved a mean C-index of 0.798 (95% CI 0.712-0.871) and a mean Brier score of 0.208 (95% CI 0.176-0.244). The calibration intercept was -0.12 (95% CI - 0.41 to 0.18) and calibration slope was 0.85 (95% CI 0.62-1.12), indicating no systematic overprediction but some risk of overconfidence in extreme predictions. For illustrative purposes only, a single 7:3 split yielded a C-index of 0.812 (95% CI 0.684-0.940), but this estimate is known to be unstable and optimistic; therefore, it is not reported as a primary finding.

An interpretable XGBoost model incorporating seven routinely available features showed good discrimination and reasonable calibration for associating with current malnutrition risk as defined by PG-SGA in PD patients. However, because several predictors (e.g., handgrip strength, triceps skinfold thickness) are components or direct reflections of nutritional status themselves, the observed performance may partly reflect conceptual overlap with the outcome rather than genuine predictive capacity. Given this conceptual overlap, the model's performance may partly reflect inherent consistency with the PG-SGA rather than genuine independent predictive capacity. The model should therefore be viewed strictly as an alternative representation of the PG-SGA-based assessment, identifying correlates of current nutritional status, rather than as a tool providing independent risk prediction. Consequently, the model's performance estimates, particularly the C-index with a 95% confidence interval whose lower bound is close to 0.7, are subject to considerable uncertainty. Given the modest sample size (events-per-variable ratio of 9.7, below the recommended minimum of 10), cross-sectional design, and the fact that three pre-specified models were tuned and compared simultaneously-an approach that carries a high risk of overfitting and chance findings-these findings should be interpreted as exploratory. Therefore, strong warnings against direct extrapolation of this model to clinical practice are warranted, and external validation in much larger cohorts is strictly required before any conclusion can be drawn regarding generalizability. The online tool is provided as an exploratory research prototype only and is not ready for clinical deployment.